Detecting Anomalies in Small Unmanned Aerial Systems via Graphical Normalizing Flows

With the increasing deployment of small unmanned aerial systems (sUASs) on various tasks, it becomes crucial to analyze and detect anomalies from their flight logs. To support research in this area, we curate Drone Log Anomaly (DLA), the first real-world time series anomaly detection dataset in the domain of sUASs, which contains 41 sUAS flight logs annotated with various types of anomalies. As anomalies tend to occur in low-density areas within a distribution, we propose graphical normalizing flows (GNF), a graph-based autoregressive deep learning model, to perform anomaly detection through density estimation. GNF contains 1) a temporal encoding module using a transformer to capture the temporal dynamics, 2) an interfeature encoding module leveraging graph representation learning on a Bayesian network to model the statistical dependencies among time series features, and 3) a density-estimation module with normalizing flows. Extensive experiments have demonstrated GNF’s superior anomaly detection power on DLA compared with state-of-the-art baselines.